RTX Spark Linux unlocks 116 GiB: 27B-class FP16 fits in 128GB unified memory for the first time
Community hands-on (Arch Linux ARM + NVIDIA 615.78): the 128GB Spark unit now exposes a 116.0 GiB unified memory pool to both OS and CUDA under Linux (GPU-usable under Windows is ~62.5GB). We added 44 Ubuntu runs: 27B-class FP16 (107GB), 31B-class Q8_0 and 30B-A3B FP16 (110GB) run for the first time — big-quant volume arriving at physical expectations on 120GB/s bandwidth.
Source: araoai
RTX Spark Linux unlocks 116 GiB: 27B-class FP16 fits in 128GB unified memory for the first time
The RTX Spark launch left a caveat: under Windows the 128GB unit is a fixed split, with only ~62.5GB GPU-usable — barely half the advertised capacity. On Oct 19 a community hands-on tore that veil off: Linux exposes a 116.0 GiB unified memory pool.
Unlock measurements (josie.lol, 2026-10-19)
- Machine: ASUS ProArt P16 128GB (RTX Spark)
- OS: Arch Linux ARM + NVIDIA 615.78.08 driver
- Result: both the OS and CUDA see a 116.0 GiB unified pool (~90.6% of the advertised 128GB; the rest is system reserve)
- CPU performance matches Windows (7-Zip within ±1.2%) — not a half-machine trade for switching OS
Against the ~62.5GB GPU-usable split under Windows, usable capacity nearly doubles.
What fits after the unlock (44 new Ubuntu runs on-site)
This batch adds 44 L2 records (Ubuntu 24.04 / NVIDIA 615.78 / llama.cpp + Ollama), all aimed at the quant tiers that fit in the 116 GiB pool for the first time:
| Config | VRAM used | Measured speed |
|---|---|---|
| Hy-Memory-30B-A3B FP16 | 110.2 GB | ~43-55 tok/s |
| gemma-4-31B-it FP16 | 110.2 GB | ~4-5 tok/s |
| Qwen3.6-27B FP16 | 107.0 GB | ~5-6 tok/s |
| Qwen3.6-35B-A3B Q8_0 | 86.6 GB | ~52-63 tok/s |
| Qwen3.6-27B / gemma-4-31B Q8_0 | 73-80 GB | 6-8 tok/s |
Three takeaways:
- The A3B family stays fluid at big quants: even at FP16 (~104GB of active weights amortized) Hy-Memory-30B-A3B still delivers 40+ tok/s on 120GB/s — sparse activation is the unified-memory device's friend
- Dense 27-31B FP16 is "can run", not "runs fast": every token reads the full FP16 weights, so 120GB/s caps out at 4-6 tok/s; Q8_0 only gets you 6-8 tok/s. Speed here waits on bandwidth, not capacity
- Beam-1.0-501B-A23B is still one notch short: IQ1_S weights run ~122GB — no room left for KV cache in the 116 GiB pool. The 501B tier remains the M5 Ultra 512GB class
What about Windows?
The fixed split stands; the 62.5GB figure is unchanged. The Linux unlock mostly proves option value: the hardware can address 116 GiB, the gap is driver and OS work. If you own one, Linux eats the full pool today. If you are on the fence, the big-quant story holds — halfway.
On-site updates
- NVIDIA RTX Spark hardware row now records the Linux unlock (116.0 GiB / driver 615.78 / ProArt P16 hands-on)
- 44 Ubuntu runs seeded against the 116 GiB threshold
- llama.cpp Q4_K_M anchors unchanged: Llama-3.1-8B at 54.7 tok/s, Qwen3.6-35B-A3B at 88.7 tok/s (aligned with StorageReview)